© 2012, IJARCSSE All Rights Reserved Page | 1

Volume 2, Issue 6, June 2012 ISSN: 2277 128X

International Journal of Advanced Research in Computer Science and Software Engineering Research Paper Available online at: www.ijarcsse.com

Fingerprint Recognition using Robust Local Features Madhuri and Richa Mishra

Department of Computer Science & Engineering,

Galgotia's College of Engineering and Technology, Greater Noida, INDIA

Abstract— There exist many human recognition techniques which are based on fingerprints. Most of these techniques use

minutiae points for fingerprint representation and matching. However, these techniques are not rotation invariant and fail

when enrolled image of a person is matched with a rotated test image. Moreover, such techniques fail when partial fingerprint

images are matched. This paper proposes a fingerprint recognition technique which uses local robust features for fingerprint

representation and matching. The technique performs well in presence of rotation and able to carry out recognition in

presence of partial fingerprints. Experiments are performed using a database of 200 images collected from 100 subjects, 2

images per subject. The technique has produced a recognition accuracy of 99.46% with an equal error rate of 0.54%.

Keywords— Biometrics, Fingerprint Recognition, Rotation and Occlusion Invariance, Partial Fingerprints

I. INTRODUCTION

Traditional Security Methods are based on things like

Passwords and PINs. However, there are problems with

these methods. For example, passwords and PINs can be

forgotten or stolen. Use of biometrics has helped in

handling these issues. Biometrics deals with the

recognition of a person using his or her biometric

characteristics [1]. There are two types of biometric

characteristics a person possesses. One is physiological

characteristics where as another is behavioral

characteristics. Physiological characteristics are unique

characteristics physically present in human body.

Examples of physiological biometric characteristics

include face, fingerprint, iris, ear etc. Behavioural

characteristics are related to behaviour of a person.

Examples of behavioral biometrics include signature,

voice, gait (waking pattern) etc. The advantage of

biometrics is that biometric identity is always carried by

a person. So there is no chance of losing or forgetting it.

Also, it is difficult to forge or steal biometric identity.

Fingerprint is one of the popular biometric trait used for

recognizing a person. Properties which make fingerprint

popular are its vide acceptability in public and ease in

collecting the fingerprint data [2,3].

Many researchers have attempted to use fingerprints

for human recognition for a long time. Most of them

make use of minutiae based approach for representation

and matching of fingerprints. Fingerprint matching based

on minutiae features is a well studied problem. These

technique often makes assumption that the two

fingerprints to be matched are of approximately same

size. However, this assumption is not valid in general.

For example, matching of partial fingerprints will not

bind by this assumption. Even two fingerprints captured

using two different scanners may have different size.

Matching of two latent fingerprints may face the same

problem. Moreover, two images with different

orientation may fail to match in minutiae based

techniques due to relative change in their minutiae

locations.

In this paper we propose a fingerprint recognition

technique which is based on local robust features. The

technique is robust in the sense that it is able to perform

person recognition in presence of rotated and partial

fingerprint images. Rest of the paper is organized as

follows. In Section II, issue with existing fingerprint

images are highlighted. Next section presents local

robust features used in the proposed technique. Section

IV presents the proposed fingerprint recognition

technique. Experimental results are presented and

analyzed in Section V. Finally paper is concluded in the

last section.

II. ISSUES WITH EXISTING FINGERPRINT

RECOGNITION TECHNIQUES

Most of the existing fingerprint techniques in

literature are based on minutiae points which are

represented using their co-ordinate locations in the

image. When test fingerprint image is rotated with

respect to enrolled image or partially available, these

techniques face problem in matching due to change in

the co-ordinate locations of the minutiae points and

perform very poorly. These two cases are discussed

below.

A. Rotated Fingerprint Matching:

An example of a rotated fingerprint image is shown

in Figure 1(b). We can see that it is difficult to match

minutiae of two images because due to rotation,

coordinate locations of all the minutiae points in Figure

1(b) with respect to Figure 1(a) are changed.

B. Partial Fingerprint Matching:

An example of partial fingerprint is given in Figure

2(b). We can see that it is difficult to match minutiae of

two images because due to missing part of the

Volume 2, Issue 6, June 2012 www.ijarcsse.com

© 2012, IJARCSSE All Rights Reserved Page | 2

fingerprint, coordinate locations of all the minutiae

points in Figure 2(b) with respect to Figure 2(a) are

changed.

Fig 1 (a) Normal Fingerprint Image, (b) Rotated Fingerprint Image

Fig 2 (a) Full Fingerprint (b) partial Fingerprint Image

Concisely, matching of rotated or partial fingerprints to

full enrolled images present in the database face several

challenges: (a) If test image is rotated, the co-ordinate

locations of minutiae points may change even with slight

rotation, (b) the number of minutiae points available in

partial fingerprints are relatively less, leading to less

discrimination power (c) co-ordinate locations of

minutiae points are also bound to change due to change

in reference point in case of partial fingerprints.

III. LOCAL ROBUST FEATURE USED IN PROPOSED

TECHNIQUE

To overcome the issues faced by minutiea based

techniques, we propose the use of local robust fetures for

fingerprint representation and matching. Among various

local features such as SIFT [6], SURF [4,5], GLOH [7]

etc available in literature, SURF (Speeded up Robust

Features) have been reported to be robust and distinctive

in representing local image information [6]. SURF is

found to be rotation-invariant interest point detector and

descriptor. It is robust with scale and illumination

changes and occlusion.

A. Key-Point detection

SURF identifies important feature points commonly

called called key-points in the image. It uses hessian

matrix for detecting key-points. For a given point in an

image I, the hessian matrix is defined as:

where , , and are filter matrices defined

as follows where gray pixels represent 0.

Key-points at different scales are detected by

considering filters at various scales. In order to localize

interest points in the image and over scales, maximum

filter in a neighborhood is implemented.

B. Computation of Descriptor Vector

In order to generate key point descriptor vector, a

region around the key-point is considered and Haar

wavelet filter responses in hohizontal ( ) and in

vertical ( ) directions are computed. These responses

are used to obtain the dominant orientation in the

circular region. Feature vectors are measured relative to

the dominant orientation resulting the generated vectors

invariant to image rotation.

A square region around each key-point is considered

and it is aligned along the direction of dominant

orientation. The square region is further divided into

sub-regions and Haar wavelet responses are

computed for each sub-region. The sum of the wavelet

responses ( and ) in horizontal and vertical

directions and of their absolute values ( and ) for

each sub-region are used as feature values. Thus, the

feature vector for sub-region is given by

SURF feature vector of a key-point is obtained by

concatenating feature vectors ( ) from all sixteen sub-

regions around the key-point resulting a vector of

elements.

Fig. 3 Flow chart of the proposed Technique

IV. STEPS INVOLVED IN PROPOSED TECHNIQUE

There are three basic modules in the proposed technique.

Various modules of the proposed technique are

discussed below. Figure 3 shows the block diagram of

the technique.

Volume 2, Issue 6, June 2012 www.ijarcsse.com

© 2012, IJARCSSE All Rights Reserved Page | 3

A. Image Acquisition

This module is used to read fingerprint images. We

collect the data using SecuGen Fingerprint scanner and

images are collected at 500 dpi.

B. Image Enhancement

In this module, image enhancement is carried out to

remove the noise from fingerprint image. We use

Gaussian smoothing filter for noise removal. We also

used average and median filter for noise removal,

however found the best result in case of Gaussian filter.

C. Feature Extraction

In the feature extraction module, features from the

enhanced fingerprint image are extracted. We have used

SURF for feature extraction. The reason behind using

SURF is that it is robust against rotation. Also, since

SURF represents image using local features, it also

works well in presence of occlusion i.e. for partial

fingerprint image.

D. Matching

In matching module, two fingerprint images are matched

with the help of extracted local features. Depending

upon the obtained matching score, two fingerprints are

declared as matched or not-matched.

V. RESULTS

Experiments have been performed on a database of 200

images collected from 100 subjects, 2 images per

subject. Few sample images from the database are shown

in Figure 4.

Fig. 4: Few samples of fingerprint images from the database

Figure 5 shows the score distribution for genuine and

imposter matches. It is clear from the figure that these

scores are quite distinguishable. This shows that the

proposed technique would be efficiently able to

differentiate between genuine and imposter matches.

Performance of the proposed technique is presented in

TABLE I. Threshold vs. FAR, FRR curves are shown in

Figure 6 whereas Receiver Operating Characteristics

(ROC) curve and accuracy curves are shown in Figures 7

and 8 respectively. TABLE II Performance of the Proposed Technique

Parameter Result

Accuracy 99.46%

Equal Error Rate (EER) 0.54 %

Fig. 5 Genuine and Imposter Match Score Distribution (100 genuine

score points against 9900 imposter score points)

(a)

(b)

Fig. 6 (a) Threshold vs. FAR, FRR plots, (b) Close View of Plots

Volume 2, Issue 6, June 2012 www.ijarcsse.com

© 2012, IJARCSSE All Rights Reserved Page | 4

(a)

(b)

Fig. 7 (a) ROC curve, (b) Close View of ROC Curve

A. Experiment to Show Rotation Invariance

All 200 images from the database are used in this

experiment. 100 images are used for enrollment and 100

for testing. To test robustness against rotation, test

images are rotated before matching. We have used

rotation angles of 50, 10

0, 15

0 and 20

0 in the experiments.

Rotated images for few angles for a subject are shown in

Figure 9. Obtained experimental results are presented in

Table II. From the table, we can observe that even after a

rotation of 100, recognition accuracy is more than 99%.

TABLE II: Performance of the Proposed Technique

in presence of Rotation

Rotation Angle Accuracy

0 99.46

5 99.22

10 99.13

15 98.15

20 94.85

B. Experiment to Show Robustness against Partial

Fingerprints

In this experiments also all 200 images from the

database are used for experimentation. 100 images are

used for enrollment while 100 for testing. To test

robustness against occlusion (partial fingerprint), test

images are partially presented during matching. Few

examples of occluded (partial) fingerprint images are

shown in Figure 10. We have experimented by

considering partial fingerprint with 5%, 10%, 15%, 20%,

25%, 30% and 40% occlusion. Experimental findings are

presented in Table III. From the table, we can observe

that even by using fingerprint with 15% occlusion,

recognition accuracy value is more than 90%. Also in

presence of 30% occlusion, recognition accuracy is

around 97%.

(a)

(b)

Fig 8 (a) Accuracy curve, (b) Close View of Accuracy Curve at

Maximum Accuracy Point

Fig. 9 Examples of Rotated Fingerprint Images

Volume 2, Issue 6, June 2012 www.ijarcsse.com

© 2012, IJARCSSE All Rights Reserved Page | 5

Fig 10 Examples of Occluded (Partial) Fingerprint Images

TABLE III Performance of the Proposed Technique

for Partial Fingerprint Images

%age Occlusion Accuracy

0 99.46

5 99.45

10 99.22

15 99.02

20 98.69

25 97.88

30 96.94

VI. CONCLUSIONS

This paper has proposed an efficient fingerprint

recognition technique which is based on local robust

features. It has used Speeded-up Robust Features

(SURF) as local robust features as it has been found to

be superior as compared to other local features in terms

of accuracy and speed. The technique has performed

well in presence of rotation and partial fingerprint

images. Experimental validation has been performed on

a database of 200 images collected from 100 subjects, 2

images per subject. The performance of the technique

has been found to be very encouraging. It has produced a

recognition accuracy of 99.46% with an equal error rate

of 0.54%.

REFERENCES

[1] A.K. Jain, A. Ross, S. Prabhakar, An introduction

to biometric recognition, IEEE Transactions on

Circuits and Systems for Video Technology,

Special Issue on Image- and Video-Based

Biometrics 14 (1), pp. 4–20, 2004.

[2] Nalini Ratha and Ruud Bolle, Automatic

Fingerprint Recognition Systems, Springer-verlag,

2003.

[3] Davide Maltoni, Dario Maio, A. K. Jain and Salil

Prabhakar, Handbook of Fingerprint Recognition,

Second Edition, Springer, 2009.

[4] Herbert Bay, Andreas Ess, Tinne Tuytelaars and

Luc Van Gool, Speeded-Up Robust Features

(SURF). Computer Vision and Image

Understanding, 110(3), pp. 346-359, 2008.

[5] Herbert Bay, Tinne Tuytelaars and Luc Van Gool,

SURF: Speeded up robust features, In Proceedings

of 9th European Conference on Computer Vision

(ECCV' 06), pp. 404-417, 2006

.

[6] David G. Lowe, Object recognition from local

scale-invariant features, Proceedings of the

International Conference on Computer Vision. 2.

pp. 1150–1157, 1999.

[7] Krystian Mikolajczyk and Cordelia Schmid, A

performance evaluation of local descriptors, IEEE

Transactions on Pattern Analysis and Machine

Intelligence, 27(10), pp. 1615--1630, 2005.